xAI: Grok 4 Fast

Processes both text and image inputs simultaneously within a 2M token context window, enabling analysis of long documents, multiple images, and extended conversations without context truncation. The model uses a unified transformer architecture that interleaves vision and language tokens, allowing it to maintain coherence across extended sequences while performing joint reasoning over heterogeneous input modalities.

Delivers state-of-the-art cost-per-token pricing while maintaining competitive performance on standard benchmarks, achieved through architectural optimizations including quantization-aware training, efficient attention mechanisms, and parameter sharing. The model is designed to minimize computational overhead during inference without sacrificing output quality, making it suitable for high-volume production workloads where cost per inference is a primary constraint.

Provides rapid text and image understanding without explicit chain-of-thought reasoning, optimized for latency-sensitive applications where response time is critical. This variant skips intermediate reasoning steps and directly generates outputs, reducing token generation overhead and wall-clock inference time while maintaining quality for straightforward tasks that don't require deep multi-step reasoning.

Generates explicit, step-by-step reasoning traces before producing final outputs, enabling transparent multi-step problem solving and verification of model reasoning. This variant allocates additional tokens to intermediate reasoning steps, allowing the model to decompose complex problems, explore multiple solution paths, and provide auditable reasoning chains that can be inspected and validated by downstream systems or human reviewers.

Exposes Grok 4 Fast through REST API endpoints (via OpenRouter or xAI) with support for streaming responses, enabling real-time token-by-token output delivery. The API implements standard OpenAI-compatible interfaces, allowing developers to integrate the model using existing client libraries and middleware without custom integration code. Streaming support enables progressive rendering of responses in user-facing applications, improving perceived latency and enabling cancellation of long-running requests.

Processes images as native inputs alongside text, enabling joint reasoning over visual and textual content. The model uses a vision encoder that converts images into token sequences, which are interleaved with text tokens in the transformer, allowing it to answer questions about images, extract information from visual content, and perform cross-modal reasoning. Supports multiple image formats and resolutions with automatic scaling to fit within the context window.